SciTabAlign
收藏资源简介:
SciTabAlign数据集是针对科学论文中的表-文对齐任务而构建的,它要求模型不仅预测陈述的真实性,还必须识别对验证陈述至关重要的表格单元格。该数据集由372个经过人工标注的样本组成,每个样本包含一个陈述、一个表格和一组被标注为支持或反驳该陈述的单元格。SciTabAlign数据集的创建过程包括对SciTab数据集中支持的陈述和反驳的陈述进行人工标注,并引入了一个包含五种模糊类型的分类法,用于处理在标注过程中遇到的边缘情况。该数据集可用于评估各种大型语言模型在科学事实验证任务中的推理和可解释性能力,并推动科学阅读工具的发展,以提高表格证据的可访问性和可解释性。
The SciTabAlign dataset is developed for the table-text alignment task in scientific papers. It mandates models to not only predict the veracity of a given statement, but also identify the table cells critical to verifying that statement. This dataset comprises 372 manually annotated samples, each containing a statement, a table, and a set of cells labeled as either supporting or refuting the statement. The construction of SciTabAlign involves manually annotating supporting and refuting statements from the original SciTab dataset, and introduces a taxonomy encompassing five types of ambiguity to handle edge cases encountered during the annotation process. This dataset can be used to evaluate the reasoning and interpretability capabilities of various Large Language Models (LLMs) in scientific fact verification tasks, and promote the development of scientific reading tools to improve the accessibility and interpretability of tabular evidence.
数据集概述
数据集基本信息
- 名称: Table-Text Alignment: Explaining Claim Verification
- 论文链接: Table-Text Alignment: Explaining Claim Verification Against Tables in Scientific Papers
数据集内容
- 数据文件: data.zip
- 输出文件: outputs.zip
任务类型
-
Claim Prediction Task
- 运行命令:
python3 run_claim.py - 评估命令:
python3 run_eval.py claim_task
- 运行命令:
-
Cell-level Evidence Selection Task
- 运行命令:
python3 run_evi.py - 评估命令:
python3 run_eval.py evi_task
- 运行命令:
结果复现
- Claim Prediction Results: 使用命令
python3 run_eval.py claim_task - Evidence Selection Results: 使用命令
python3 run_eval.py evi_task




